The headline landed in my feed at 7:14 a.m. Saigon time, which is the hour when the US close finishes printing and every aggregator pushes its morning summary into Southeast Asia. “Howard Marks is not convinced the AI boom is a bubble.” Four facts in the body. Zero direct quotations. No memo link. No byline I could trace. I read it twice, then opened my capex spreadsheet, because that is what I do when a headline makes an emotional claim and a financial claim in the same sentence.
Here is the arithmetic the headline left out. The four largest North American cloud operators have been guiding, across the last two reporting cycles I tracked, toward a combined annual capital expenditure figure in the region of $300 billion to $350 billion. The two most commercially successful frontier model companies were running annualized revenue in the low tens of billions combined. Even if you generously fold in the AI-attributable slice of cloud revenue, the ratio between money going into the ground and money coming out of the product has no recent precedent outside of infrastructure manias.
That is not a disagreement about whether AI works. It is a disagreement about a denominator. And in a bear market, the denominator is the only thing that decides whether you survive long enough to be right.
Howard Marks is a credit investor. His bubble radar points at the debt stack, not the price-to-earnings ratio. This matters more than anything else in the article, and the article never says it. Oaktree built its reputation buying distressed debt in cycles where the equity holders were wiped out and the senior lenders collected. When a man with that balance sheet of experience talks about bubbles, he is not talking about whether a stock is expensive. He is talking about who is holding the paper when the cash flow stops.
His framework, as I have read it across two decades of memos, has never been technology-dependent. It is price-dependent. The test is not whether the asset rose a lot or the narrative got loud. The test is threefold: is psychology overheated, has price already fully discounted the optimistic scenario, and have the incremental buyers been exhausted. The same technology is an opportunity at one price and a bubble at another. That is not a hedge. That is the whole argument.
The headline stripped the argument out and kept the verdict. “Not convinced the boom is a bubble” reads, in the mind of anyone scrolling, as an authority figure issuing a denial. The body summary, by contrast, still contained the warning about overvalued valuations. Those two statements have nearly opposite operational meanings. This is a framing effect, not a factual error, which is precisely why no fact-checking layer will ever catch it. There is no false statement to correct. There is only a condition that got dropped between the memo and the push notification.
I don't trade a headline I haven't read past.
The sentence that went missing is almost certainly some version of: the question is not whether AI is important, but what you paid for it. Remove that clause and a risk warning becomes a mood stabilizer. In a market where retail participation is high and liquidity is thin, a mood stabilizer has measurable price impact. That is not a small editorial sin. It is a distribution mechanism for mispricing.
I have seen this movie at the contract level. In late 2017 I spent three weeks auditing the ERC-20 distribution logic of a mid-tier ICO that was raising twelve million dollars. The whitepaper was flawless. The tokenomics diagram was beautiful. The bug was an integer overflow in the mint path that would have allowed unlimited supply. Nobody had read the code, because everyone was reading the deck. The lesson I took from that quarter and have never dropped: the risk always lives in the artifact nobody opens. For a token, it is the contract. For an AI equity, it is the depreciation schedule and the debt covenant.
So let me open the artifact.
The most systematically ignored number in this entire debate is the useful life a cloud operator assigns to its accelerators. Five to six years is the common assumption. The economic life, given the cadence of silicon refreshes and the way inference workloads migrate to newer, cheaper, more efficient parts, is plausibly two to three years. This is not fraud. It is a difference of assumption between people who are paid to make assumptions. But the arithmetic is brutal.
Take a fleet large enough to absorb one hundred billion dollars of accelerator spend. Straight-line over six years is roughly $16.7 billion of annual depreciation. Over three years it is $33.3 billion. The delta is $16.6 billion a year of pre-tax earnings, moved by a single accounting choice, disclosed in a footnote. That delta is on the order of the entire annualized revenue of the leading frontier model company. You are not arguing about a rounding error. You are arguing about whether the profit pool reported by the most valuable companies in the world is real or borrowed from a future write-down.
And because useful-life assumptions are disclosed and must be revised when management's judgment changes, they are a leading indicator of profit quality. When a company shortens its assumed life, it is not confessing a mistake. It is repricing reality. Watch for that revision. It will arrive quietly, in a 10-Q, on a Tuesday, months before anyone writes the retrospective.
The second artifact is the financing structure, and this is where Marks's actual expertise would bite. A meaningful share of this build-out is not being funded from operating cash flow alone. It runs through special purpose vehicles for data center construction, through joint ventures that keep obligations off the primary balance sheet, through private credit funds hungry for yield in a high-rate world, and through vendor financing — the chip supplier invests in the customer, the customer commits to buy chips, and the investment sits on one balance sheet as an asset while the purchase sits on the other as revenue.
None of that is illegal. All of it is a structure. We ran this experiment in 2000 with Lucent and Nortel financing the competitive local exchanges that were buying their equipment. The equipment revenue was real when it was booked. The cash never came back. Vendor financing does not create demand. It pulls demand forward and hands the receivable to someone who did not underwrite it. When the music stops, the first price to break is not the equity. It is the credit spread on the paper nobody was watching.
The third artifact is concentration. The ten largest names in the S&P 500 now carry something on the order of a third or more of the index weight, a level of single-theme concentration that exceeds the 2000 episode. That converts an AI repricing from a sector event into an index event. It transmits through passive flows, through pension allocations, through household balance sheets, and eventually into consumption. This is the channel that turns a technology debate into a macro problem, and it is the channel the headline article did not mention once.
The fourth artifact is game theory. Every hyperscaler knows the return on this capex is unproven. Every hyperscaler also knows that unilateral disarmament means losing share in the next workload migration, which is a permanent loss, not a cyclical one. So individual rationality produces collective overinvestment. This is a prisoner's dilemma with a capex line, and prisoner's dilemmas do not resolve through rational reappraisal. They resolve through external constraint. Power, financing cost, or demand falsification. Nothing else closes the position.
Power is the constraint I find most interesting because it works in both directions. Interconnection queues, transformer lead times, and the signing cadence on nuclear power purchase agreements put a physical ceiling on how fast capacity can be built. That is a brake on the bubble inflating without limit. It is also the strongest argument that compute remains scarce, which supports pricing. Both readings are defensible, which means both must be verified separately rather than asserted together. The brake and the scarcity are the same fact wearing different clothes.
Now the part I actually know something about, because I lived it in 2020.
During DeFi summer I wrote a Python bot that watched Uniswap and SushiSwap pools and fired arbitrage trades when the implied prices diverged. Five hundred plus executions, forty-five thousand dollars of profit, and one lesson that has never left me: a subsidy looks exactly like demand on a chart. Total value locked went vertical. It looked like adoption. It was an emissions schedule. When the liquidity mining incentives decayed, the TVL left within days, because it had never been capital. It had been mercenary capital renting a yield.
I want to say this carefully, because the analogy is not perfect and I don't want to overdraw it. AI capital expenditure is the emissions schedule of the largest liquidity mining program in history. The difference is that the subsidy here is being paid by the protocol itself, out of operating cash flow and debt, to acquire a position in a market that may or may not exist at the size the capex implies. The question an analyst should be asking is the DeFi question: what is the organic TVL? What revenue survives if the emissions stop?
That question is hard to answer because almost nobody discloses revenue quality. A large share of enterprise AI spend may be budget substitution — software and cloud line items relabeled — rather than net-new productivity spending. If that is the case, the aggregate demand increment is being overstated, and the capex is being sized against a denominator that was already counted once. This is the single most important unknown in the entire debate, and it is exactly the unknown that opinion journalism cannot resolve, because resolving it requires primary research rather than a quotation.
Last year I built something that gave me a partial answer. I wrote an autonomous agent that negotiated data access fees over Ethereum, holding a testnet wallet with ten thousand dollars in it. The agent did not care about brands. It routed on price and latency. It paid in fractions of a cent per call, and it rejected any vendor whose marginal price exceeded the value of the task it was trying to complete. That last behavior is the one that reshaped how I model this market.
Agentic demand is metered demand, and metered demand has a hard price ceiling set by the value of the marginal task. Seat-based software pricing had inertia: once a workflow lived inside a tool, the buyer stopped shopping. An agent shops on every single call. That means the application layer has far less pricing power than the SaaS multiples embedded in current valuations assume. And the infrastructure layer's pricing power depends entirely on scarcity — the very scarcity that three hundred billion dollars of annual capex is engineered to destroy.
Run the if-then. If inference cost per token falls tenfold while agent call volume rises tenfold, revenue is flat, gross profit is roughly flat, and the capex was sized for a volume curve that arrived but did not monetize. That is not a doomsday scenario. It is the base case of every efficiency curve in computing history. The bear case does not require AI to fail. It only requires AI to succeed at exactly the rate that competition forces prices down.
And here is where this stops being a story about Nvidia and starts being a story about the assets I actually underwrite. The crypto transmission channel for this narrative is direct and it is levered. Agent tokens, decentralized physical infrastructure networks selling fractional GPU capacity, tokenized compute marketplaces — every one of them is a higher-beta, thinner-liquidity expression of the same underlying thesis. When risk appetite compresses, these are the first to bleed, because their marginal holder is the marginal risk-appetite buyer in the entire market. They do not have buybacks. They have a chart and a Telegram channel.
I have watched the compute-marketplace pitch deck at least a dozen times this cycle, and it has the same structural flaw I have been writing about in the Layer 2 space for two years. Dozens of venues, all competing for the same small pool of real demand, each one calling fragmentation capacity. Slicing scarce liquidity into fragments is not scaling. It is the appearance of scaling, and it is manufactured by people who need a new product to sell. The decentralized compute market is running that playbook in real time, one aggregator at a time.
So where does the diffusion of this debate leave us? When the bubble argument migrates from technology media to crypto media, and from growth investors to value investors, that migration is itself a data point. Historically it marks the middle-to-late phase of a narrative cycle, not the beginning. And I weight value-investor commentary accordingly, because value managers tend to speak publicly about a mania at the moment they must explain why they missed it, not at the moment they are positioning for it. That is not cynicism. It is an incentive read.
Now the part I think both sides are getting wrong.
The consensus bear case assumes a 2000-style event: a vertical run, a single trigger, a capitulation candle, a clean entry at the bottom. I don't think that is the shape of this. The physical constraints — grid interconnection, transformer supply, power purchase agreement lead times — act as a brake on the upswing, and a brake on the upswing is also a brake on the collapse. What you get instead of a crash is attrition. Capacity arrives in tranches. Revenue grows slower than the depreciation attached to it. Margins compress for six or eight quarters while everyone insists the inflection is next quarter. The multiple does not collapse. It bleeds.
A slow mispricing is more dangerous to allocators than a crash, because it never gives you the capitulation candle that makes buying feel safe. The crash is honest. It marks the price and lets you act. Attrition just grinds the return out of a position while the story stays intact, and the story staying intact is what keeps you in it. That is the blind spot on the bear side.
The blind spot on the bull side is reflexivity. If “AI is a bubble” becomes the consensus, capital withdraws ahead of the fundamentals, and the withdrawal genuinely slows deployment, which validates the thesis that caused the withdrawal. The narrative manufactures the outcome it predicted. That cuts both ways: a bear case that becomes too popular destroys its own premise by clearing out the marginal seller and handing the survivors better unit economics. I have no clean way to trade that. Nobody does. It is the reason I size positions rather than predict prices.
But the contrarian conclusion I will actually commit to is this: the bubble may not be in the equities at all. It may be in the narrative layer, and in the token layer that levered it. Equity holders own cash flows, buybacks, and a legal claim. They can be wrong about the multiple and still hold a productive asset. A token holder owns a story with a liquidity profile. When the story compresses, there is no residual claim underneath it. The equities can be expensive and survive. Most of the AI-adjacent tokens cannot.
Which brings me back to the four facts and the missing quotation. The article is not useful as information. It is useful as a timestamp. It records the moment the bubble debate reached the audience with the highest beta and the least patience. Treat it as a lead to be verified, not a conclusion to be traded.
So here is what I am watching, and I would suggest you watch the same things. Accelerator rental spot prices first — a sustained decline in hourly rates for the previous-generation parts is the earliest available signal that demand is softening, because it clears before the contract renegotiations do. Then data center private credit and ABS spreads, which will widen before any equity cracks, because credit reprices on uncertainty while equity reprices on earnings. Then useful-life revisions in the quarterly filings, which tell you what management actually believes about the economic life of the assets it just bought. Then, quarterly, the ratio of capex guidance to AI revenue run-rate. If that ratio stops widening, the story is working. If it widens while the multiple holds, you are being paid in narrative.
None of these require an opinion about whether AI is transformative. They only require arithmetic and a calendar.
Arbitrage is just geometry disguised as finance. The distance between what a headline says and what a source said is a spread like any other — narrow, exploitable, and closed within days by anyone willing to read the original. The distance between capex and the revenue it is supposed to produce is a wider spread, and it closes on a schedule nobody controls. The only question worth asking is which side of that spread you are standing on when the emissions stop.